How Many Photos Do You Need for an AI Video of Yourself?
The models accept up to nine. Most people supply one, which is the single most common reason generated faces do not look like the person. What to shoot and why the count matters.
Short answer: four is the practical minimum, nine is the ceiling on most models, and one is why your result did not look like you.
What the models actually accept
| Model | Max references | $/second |
|---|---|---|
| HappyHorse 1.1 Reference to Video | 9 | $0.14 |
| MiniMax H3 Reference to Video | 9+ images, plus reference clips | $0.26 |
| Kling O3 Pro Reference to Video | 4 | $0.14 |
Prices checked against the live endpoints on 3 August 2026.
On HappyHorse the references are not merely uploaded — they are cited in the prompt by position, as character1 through character9, matching the order you supply them. So the prompt does not say "a woman in a kitchen"; it names character1 and describes what they do. That mechanic is why supplying more angles helps: the model is matching a named subject across viewpoints, not averaging a set of pictures.
Why one photo fails
A single frontal selfie tells the model what you look like from exactly one direction. The moment your head turns — and in any clip with movement it will — the model has to invent the rest of your face. What it invents is plausible and is not you.
This shows up as the specific complaint people report as "it looks like me but isn't". It is not a quality ceiling. It is missing input.
What to shoot
Four minimum:
- Frontal, neutral expression
- Three-quarter left
- Three-quarter right
- Profile
Add if you can:
- Slight upward angle
- Slight downward angle
7–9. Full-body frames, if the clip shows your body
Conditions, which matter as much as the count:
- Same lighting across all of them — even and frontal, no heavy shadow across one side.
- Whole head in frame. No crops at the forehead or chin.
- No sunglasses, no heavy filters, no motion blur.
- The wardrobe you want in the video. Reference models read clothing and build from your photos, not from any source clip.
- Sharp. HappyHorse requires a shortest side of at least 400px and recommends 720p or better; meeting a minimum is not the same as being good.
Four good references beat nine bad ones. Nine good ones beat four.
Where the count matters most
Dance and motion transfer. Movement rotates your head far more than a talking head does, so angle coverage is doing real work. If you are using Kling 3.0 Pro Motion Control, also set character_orientation to video — on image the face-binding element silently does nothing regardless of how many photos you supplied.
Full-body shots. If the body is in frame, supply full-body references. Otherwise the model invents a build, and in fitness or fashion content that is exactly what viewers notice.
Anything longer than eight seconds. Drift compounds with duration, and a rich reference set slows it down.
What more photos will not fix
Wrong model family. If the job is copying a specific clip's choreography, no number of reference photos helps — reference models build new shots and never ingest motion from a source video. That is a routing decision, and it is the most expensive mistake available in this category.